Risk calculator for predicting postoperative pneumonia after gastroenterological surgery based on a national Japanese database

Abstract Background The aim of the present study was to develop a risk calculator predictive of postoperative pneumonia in patients undergoing gastroenterological surgery. Methods We analyzed data from 382 124 patients undergoing eight main gastroenterological surgeries between 2011 and 2013 using the National Clinical Database in Japan. A risk model was developed using multivariate logistic regression analysis with patient data from 2011 to 2012 (n = 247 604) and validated using data from 2013 (n = 134 520). Results Pneumonia was observed in 11 105 patients (2.9%). After the input of significant primary disease and surgical procedures, 18 patient characteristics including gender, chronic obstructive pulmonary disease, sepsis, and need for any assistance in the activities of daily living, six laboratory parameters, and two intraoperative factors were used for risk calculation. Area under the receiver‐operating characteristic curve was 0.822 (95% confidence interval, 0.817‐0.826) in the derivation group and 0.826 (0.819‐0.832) in the validation group. Conclusion The risk calculator accurately predicted the occurrence of pneumonia following gastroenterological surgery.

reported that postoperative pneumonia increased the risk of inhospital mortality ninefold, resulting in a mean increase of 11 days in length of hospital stay and a $28 000 US dollar increase in total hospital costs per patient. A better understanding of which patients are at an increased risk of postoperative pneumonia is important to help prioritize the introduction of enhanced preventive interventions. 6 In a study of pulmonary complications specific to pneumonia, Kinlin  Pneumonia-prevention programs have been successfully implemented not only in intensive care unit (ICU) settings but also in surgical wards. 6,17 The program in surgical wards consists of ambulation, breathing exercises, oral care, and bedhead elevation. 6 However, the pneumonia-prevention program has not been implemented routinely in patients undergoing gastroenterological surgery because of the lack of a distribution of the standard approach in non-critical care settings and the limited number of medical staff in Japan. In addition, a concerted effort is needed to maintain compliance with the program. To become a widely adopted program, identification of high-risk patients for postoperative pneumonia is required.
The aim of the present study was to develop and validate a risk calculator for predicting postoperative pneumonia after gastroenterological surgery. We used data from the gastroenterological section of the National Clinical Database (NCD) of Japan, which was established in April 2010, with 10 surgical subspecialty societies on the board of the Japan Surgical Society. 18 The NCD collaborates with the American College of Surgeons' National Surgical Quality Improvement Program (ACS-NSQIP), 19 which shares a similar goal of developing a standardized surgical database for quality improvement. 20

| Data set
The NCD is a nationwide project administered in conjunction with the board certification system for surgery in Japan. Data were extracted from the 2011-2013 NCD data files. There were 2158 participant hospitals. The NCD continuously recruits individuals who approve the data, and members of various departments are in charge of cases and data entry officers, and a web-based data management system ensures data traceability. The staff also validate data consistency through random inspections of the institutions. Surgical cases from each department were registered in the gastroenterological surgery section of the NCD, which required detailed input of eight selected main procedures. All variables and definitions, as well as the inclusion criteria for the NCD, are accessible on the NCD web site (http:// www.ncd.or.jp/). The NCD supports an E-learning system to ensure consistent data entry.

| Patients
Inclusion criteria for the study were patients who underwent the following operations: (i) esophagectomy; (ii) total gastrectomy, (iii) distal gastrectomy; (iv) right colectomy; (v) low anterior resection; (vi) hepatectomy with >1 segment except for the lateral segment; (vii) pancreaticoduodenectomy; and (viii) surgery for acute diffuse peritonitis. Patients agreed for their data to be included in the research projects by using presumed consent with an opt-out through the web page and/or a notice from each hospital. The NCD project was approved by the Japan Surgical Society Ethics Committee on November 2010. Patients who declined to have their records entered into the NCD were excluded from our analysis. Records with missing data on patient age, gender, or the occurrence of postoperative pneumonia were excluded. In the 2011-12 dataset for the development of the risk calculator, 247 604 records were used; in the 2013 dataset for the validation of the model, 134 520 records were used.

| Outcome measures
Primary outcome of interest was postoperative pneumonia.
Postoperative pneumonia was defined as pneumonia occurring within 30 days post-surgery in patients with no evidence of pneumonia preoperatively.
The registry defines postoperative pneumonia as having met one of the following conditions: c. Isolation of pathogen from specimen obtained by transtracheal aspirate, bronchial brushing, or biopsy.  In addition, we evaluated 30-day mortality and operative mortality in patients with pneumonia. The latter was defined within the index hospitalization period, regardless of the length of hospital stay (up to 90 days), as well as any death after discharge, within 30 days of surgery. Surgical site-related morbidities that occurred within 30 days of surgery (superficial/deep incisional SSI, organ space SSI, wound dehiscence, and anastomotic leak) were also assessed.

| Preoperative and intraoperative variables
NCD variables including patient demographics, pre-existing comorbidities, preoperative laboratory values and perioperative data are almost identical to those applied in ACS-NSQIP. 21 In addition to these, intraoperative factors such as prolonged surgery and severe blood loss were also evaluated. The definition of prolonged surgery is an operation lasting over the 75th percentile of the distribution of operation time for a specific category of procedures. The definition of severe blood loss is blood loss that is over the 75th percentile of the distribution for a specific category of procedures. Logistic regression models were constructed using stepwise selection of the predictors. Discriminatory ability of the prediction rule in the derivation group was quantified using the area under the receiver-operating characteristic (ROC) curve. Model calibration was examined by comparing the observed and predicted means with 10 equally sized subgroups, which were arranged in order of increasing patient risk. Model validation applied the developed model to estimate pneumonia probabilities for all patients in the 2013 dataset. A value of P < 0.05 was considered statistically significant.  Table 1.

| RE SULTS
The highest incidence of postoperative pneumonia was observed in esophagectomy (13.7%). Surgery for acute diffuse peritonitis was next highest, with an incidence of 7.5%. Gastrectomy (total/distal), pancreaticoduodenectomy and hepatectomy had incidences of 2.5% (3.6/1.9), 2.5% and 2.2%, respectively. The lowest incidence of postoperative pneumonia was observed in low anterior resection (0.9%) and right hemicolectomy (1.6%). Univariate analysis for predictors of postoperative pneumonia is shown in Table 2. We identified 38 independent predictors of postoperative pneumonia, and the multivariate logistic regression model with odds ratios (OR), 95% confidence intervals (CI) and β coefficient is shown in Table 3. After the input of significant primary disease and surgical procedures for a given patient (shown in Table 4), characteristics of 18 patients including gender, chronic obstructive pulmonary disease, sepsis and need for any assistance in the activities of daily living (ADL), six preoperative laboratory data and two intraoperative factors were used for risk calculation (Table 5). The scoring system for the postoperative pneumonia risk models based on the logistic regression equation was as follows: where β i is the coefficient of the variable X i in the logistic regression equation provided in Table 4. If a categorical risk factor is present, the X i value is 1 (0 if it is absent). For the age categories,

| D ISCUSS I ON
Estimations of the incidence of postoperative pneumonia vary widely in the literature, from 0.6% to 17.5%. [2][3][4][5][6][7][8][9][10][11][12][13][14][15]22 The variability is primarily because of the diagnosis criteria used and differing surgical procedures. In the present study, the overall incidence of postoperative pneumonia within 30 days of gastroenterological surgery was 2.9%, which is similar to the incidence reported by Yang et al 10  They analyzed all patients with a discharge diagnosis of pneumonia, including late-onset pneumonia.
In the present study, the highest incidence of postoperative pneumonia was observed after esophagectomy (13.7%) and the lowest incidence was observed after low anterior resection and right hemicolectomy (0.9% and 1.6%, respectively). Yang et al 10 described similar incidences as follows; esophagectomy 16.2%; gastrectomy 6.4%; pancreatectomy 4.8%; hepatectomy 3.3%; and colectomy/ proctectomy 2.4%. In Japan, transthoracic excision of the esophagus with extended lymph node dissection is the standard procedure for patients with esophageal cancer. In one report from Japan, the incidence of pneumonia was 8.7% in patients who underwent subtotal esophagectomy. 14  consistently identified as an independent risk factor for postoperative pulmonary complications. 13,16,23 In the multivariate logistic regression analysis, a high β coefficient was observed in males, patients with chronic obstructive pulmonary disease (COPD), patients with a need for any assistance in   F I G U R E 1 Approximated predicted risk for postoperative pneumonia. Risk for postoperative pneumonia is approximated from a nomogram of predicted probability vs calculated total risk score. β i is the coefficient of the variable X i in the logistic regression equation provided in Table 4 teaching hospitals. We developed the risk model limited to patients undergoing gastroenterological surgery. Previous studies included patients undergoing a wide range of non-cardiac operations. 5,7,8 Most of the previous studies have included a wide array of postoperative pulmonary complications such as atelectasis, pulmonary embolism and respiratory failure. 10,13,16,22 Varying definitions of postoperative pulmonary complications cause variability in risk estimates. We evaluated preoperative risk factors as well as intraoperative factors (prolonged surgery and severe hemorrhage).
Several studies have identified that these two intraoperative factors were independently related to postoperative pneumonia 7,21 or respiratory complication. 15,24 Despite the many strengths of our study, there are several

| CON CLUS IONS
We developed a risk calculator for predicting postoperative pneumonia after gastroenterological surgery using a large dataset from the NCD in Japan. Performance of the model was good in terms of discrimination and calibration, and the model was validated using datasets submitted to the NCD in different years. In patients who were assessed as high risk for postoperative pneumonia, enhanced preventive interventions should be considered. This risk model is also useful in counseling and for obtaining informed consent from patients.

ACK N OWLED G EM ENTS
The authors would like to thank all data managers and hospitals participating in the NCD project for their great efforts in entering the data, and also thank the members and the working members of the Japanese Society of Gastroenterological Surgery (JSGS) database committee.

D I SCLOS U R E
Conflicts of Interest: Authors declare no conflicts of interest for this article.